ByteDance Launches Doubao-1.5-pro Model with 50% Gross Margin

On January 22, ByteDance released the Doubao-1.5-pro large language model, boasting significant performance enhancements over its predecessor. Benchmark tests indicate that the new model surpasses GPT-4o in multiple categories while drastically reducing inference costs.

According to ByteDance, the Doubao-1.5-pro model achieves high efficiency through the use of smaller activation parameters during pre-training, keeping training costs low without compromising performance. The model employs a large-scale sparse MoE (Mixture of Experts) architecture, delivering the equivalent of a dense model with seven times the activation parameters, far surpassing the industry standard efficiency of approximately three times leverage for MoE models.

Pricing information available on Volcano Engine, ByteDance’s cloud services platform, shows competitive rates for the Doubao models. The Doubao-1.5-pro-32k is priced at CN¥0.8 (USD0.11) per million tokens for input, ¥0.16 for cached input, and ¥2 for output. The lightweight version, Doubao-1.5-lite-32k, costs ¥0.3 per million tokens for input, ¥0.06 for cached input, and ¥0.6 for output—positioning both models as the most affordable options in their class.

SEE ALSO: The Doubao App Has Been Updated with Realtime Voice Call Feature

Despite its competitive pricing, the Doubao-1.5-pro model’s optimized inference costs have led to significantly higher gross margins. According to a report by Chinese media outlet Jiemian News, an insider close to Volcano Engine revealed that while previous Doubao APIs achieved reasonable margins, the API for Doubao-1.5-pro now boasts a gross margin of 50%.

In an earlier interview, Volcano Engine President Tan Dai addressed concerns about profitability in light of model price reductions. He emphasized that achieving large-scale usage is essential for refining models and reducing per-unit inference costs. “Price cuts for large models should not be limited to lightweight versions. Core and cutting-edge models also need to be affordable to meet complex enterprise demands, validate their value, and catalyze innovation beyond current product and organizational frameworks,” he noted.

According to ByteDance, Doubao has secured partnerships with 80% of major car brands and has been integrated into numerous smart devices, including smartphones and PCs, with approximately 300 million devices now connected. Over six months, calls to the Doubao model from smart terminals increased by 100-fold.

Despite its rapid adoption, ByteDance’s large-model business remains unprofitable due to substantial R&D investments. The insider noted that long-term profitability hinges on scaling model usage further to amortize these costs over time. ByteDance has yet to comment on the matter.

Report